Mastering the game of Stratego with model-free multiagent reinforcement learning

ABSTRACT

Imperfect information games, where players have hidden information, represent a significant challenge for artificial intelligence. Stratego is a complex, imperfect-information board game with an enormous state space and exceptionally long horizons, requiring both long-term strategic planning and complex bluffing mechanics. We present DeepNash, an autonomous agent capable of learning to play Stratego from scratch to a human expert level. Instead of relying on traditional search techniques, DeepNash uses a model-free multiagent reinforcement learning algorithm based on Regularized Nash Dynamics. DeepNash achieves a top-3 ranking among all human players on the Gravon games platform, demonstrating a profound ability to execute complex deception and risk-management strategies.